How it works

One loop: capture intent → set criteria → agent implements & records → verify in isolation → traceable, done.

  1. 1

    Capture

    Everything lands on the canvas

    A scratchpad is the front door. Paste a stack trace, a customer complaint, a shower-thought feature — anything. A local model reads it and files it as a todo, bug, knowledge-base entry, or use case, anchored to the right project. No forms to fill, no folders to maintain.

  2. 2

    Define

    Say what 'done' means

    Give a piece of work acceptance criteria and the commands that prove it — your test, lint, type, and scan commands. CodeDistill detects your language and pre-fills sensible defaults, so intent becomes something a machine can actually check.

  3. 3

    Implement

    The agent records as it works

    When an agent makes a change, it records what it did against the intent that requested it. The reasoning lives with the code and the commit — not in a chat log you'll never reopen.

  4. 4

    Verify

    Checked at the real commit

    CodeDistill runs your verification commands against the recorded commit — in isolation, on demand. A criterion is satisfied only when the check actually passes. Green is earned, not assumed.

  5. 5

    Trace

    Every change has a paper trail

    From any change you can walk back to the intent that asked for it, the criteria it had to meet, and whether it met them. Risky work surfaces for review instead of slipping through.

It all runs on your machine — a local model via Ollama, storage in a single SQLite file. No cloud, no code leaving the building.

Get CodeDistill